Russell Carleton on day/night splits.
… The practice of using split statistics (home-road, day-night, left-right, April-May) is taking a stat that is already a little unreliable and cutting its underlying sample size in half (or more!)… which makes it an even more unreliable measurement. Cut the sample size to 300 PA, and AVG has a split half of .328, OBP is at .596, and SLG is at .634. Cutting the sample size in half takes something that was already teetering on the edge of respectability and makes it a worse measurement. This leads to people saying silly things like swearing up and down that there’s something about playing in the sunlight that makes Chase Headley hit like Chase Utley. So, given that I just spent a few paragraphs trashing the use of splits, why am I about to use them? Because, unlike some other Robin Hoods, I can speak with an English accent.
….
The next question is whether these differences are reliable from year to year. That is, if Johnny Damon swung three percent more during the day than at night in 2009, then would we find (roughly) the same split going back a few years? To answer this question, I found the splits for all players who met the 50 PA both at night and during the day in a season from 2006-2009. I found the intra-class correlation (AR(1) covariance matrix, for the initiated) for the change rates for each of these measures over the four years. For those who aren’t familiar, intra-class correlation is kinda like a year-to-year correlation (and you can read it the same way)—it’s just that it incorporates more than two data points.
And here’s where things fall apart. The splits on both measures fell below an ICC of .05. To put that in some perspective, BABIP for pitchers has a better ICC. So, if a player makes better contact during the day than at night in 2007, that’s interesting as a historical truth, but it has almost no predictive power as to what’s going to happen in 2008. Taylor Teagarden and Darnell McDonald are not really vampires. They just happened to have a season in 2009 where they had some weird splits. It was chance.
These findings call into question the use of platoon splits for anything other than as fodder for factoids. If, given stats that are super-reliable (as swing and contact are) that can be linked directly to the variable under study (day vs. night) by logic, there’s no reliably replicable split difference, then what hope is there for much less reliable measures which involve a great deal of luck? …
